PolicyDiff
Upload a client proposal and the carrier-issued policy, and get a line-by-line coverage discrepancy report with page citations in minutes instead of a CSR cross-reading 40 pages.
One product, one founder, thirty days. Everything below is built to be executed inside the contest window — the official rules still decide what counts.
Decide in 20 seconds
The facts that decide whether this is worth thirty days of yours.
Principal, operations manager, or senior CSR at an independent US or Canadian property and casualty agency with roughly 3 to 25 staff, at the moment a carrier-issued policy arrives and must be checked against the proposal.
Operational simplicity 3.6/10
Remove regulated data, human review steps and custom infrastructure; keep the product self-serve and stateless on managed services.
Why this could work in 30 days: one founder can ship the MVP in 5 days; the plan targets a first sale on day 12; 7 sales at $149 clears $1,000 of collected revenue; strongest signal is speed to first revenue at 7.8/10.
Score and projections come from the generated report and are not guarantees. The $100K Hackathon is run by Ship or Die — read the official rules before you build.
Your first 24 hours
13 concrete actions. Tick them off as you go — nothing is saved or sent.
Product opportunity and positioning
Target customer
Principal, operations manager, or senior CSR at an independent US or Canadian property and casualty agency with roughly 3 to 25 staff, at the moment a carrier-issued policy arrives and must be checked against the proposal.
Why now
Managed vision-language document parsing is finally good enough on dense carrier PDFs, and a solo founder can ship the finished artifact in under a week without any carrier approval, integration or procurement.
Problem
Policy checking is a manual cross-read of limits, deductibles, named insureds, effective dates and endorsement form numbers, page by page; misses surface months later as rework, lost trust and E&O exposure.
Solution
Two PDFs in, a normalized field table out, a deterministic diff that classifies every field matched, mismatched, missing or extra, and a PDF/CSV discrepancy report with page citations, source crops, reviewer initials and a read-only share link.
PolicyDiff — 30-Day Launch Report
Contest window: Nov 16 13:00 UTC to Dec 16 13:00 UTC, 2026. One founder. Money collected inside the window only.
1. Product opportunity and positioning
Independent P&C agencies with 3 to 25 staff run a policy check every time a carrier issues a policy. A CSR reads the issued policy against the proposal, field by field, confirming limits, deductibles, named insureds, effective dates and endorsement form numbers. It takes 30 to 90 minutes, it is the least popular hour of the day, and misses surface months later as rework and E&O exposure.
PolicyDiff performs the cross-read. The CSR uploads two PDFs, the proposal and the issued policy, and receives a discrepancy report: every field classified matched, mismatched, missing on policy or extra on policy, each flag carrying a page number and a cropped source image, exported as PDF and CSV with a reviewer-initials sign-off line, plus a read-only share link.
The positioning sentence is the sale: the CSR verifies, they do not transcribe. The buyer is the principal or operations manager who signs the card; the user is the senior CSR who gets the hour back. Alternatives are outsourced checking vendors (batch, opaque, expensive) or the status quo of unpaid CSR overtime. Nothing must be integrated, approved or migrated for the first policy to be checked.
The wedge is deliberately narrow: personal and small-commercial lines, one proposal against one policy, one exported artifact. No carrier form library, no coverage opinions, no agency management system sync. This is not a parsing API and not a comparison matrix for hardware engineers; it is an exception report for insurance retail.
Given a C-band JEV score of 61.6, the concept's strength is speed to first revenue and reachable buyers, and its weakness is operational simplicity. The plan below answers that weakness by keeping every flag reviewer-confirmed, keeping delivery stateless on managed services, and refusing custom infrastructure.
2. Smallest sellable MVP
A working upload page, an extraction pipeline, a deterministic diff, and an emailed report. The minimum sellable unit is: one proposal PDF plus one matching issued policy PDF in, one citation-backed discrepancy report out, delivered within 24 hours, priced at $149 per month for up to 25 checks.
In scope for day 5: page-level parse of text-based and scanned PDFs; extraction of policy number, dates, named insured, coverage lines, limits, deductibles and form numbers; normalization of currency, date formats and name word order; four-way classification per field; page citation and crop per flag; PDF and CSV export with initials; read-only share link; editable override and re-export; Stripe Checkout for the monthly plan and the $39 five-check pack.
Out of scope: agency management system integrations, ISO or carrier form libraries, coverage interpretation, commissions, multi-user roles, SSO, claims, renewals, certificates of insurance, and any promise of full automation.
Demo flow: upload proposal, upload policy, wait 90 seconds, review flagged fields, click any flag to see the source crop, correct one wrong extraction, stamp initials, download PDF and CSV, copy the share link into the client delivery email.
3. Technical implementation plan
Next.js on Vercel handles upload, review UI and report rendering. Supabase stores job rows in Postgres and both source PDFs plus generated crops in object storage. A managed layout-aware vision-language OCR endpoint extracts fields into a strict JSON schema per document type; schema validation failures are surfaced as low-confidence fields rather than guessed. pdf.js renders pages for crop generation, pdf-lib assembles the final report.
Ordering: schema first, then parse, then diff, then report, then payment. A separate prompt and schema per document family (proposal, carrier declarations page, endorsement schedule) is more reliable than one general prompt. The diff engine is ordinary code, not an LLM: exact and normalized comparison, unit and currency coercion, name word-order tolerance, and a fuzzy threshold on entity names to avoid false positives.
The single largest technical risk is scanned or poorly tagged carrier PDFs returning a wrong limit or a missed endorsement, which destroys trust on the first report. Mitigation: confidence thresholding, mandatory source crop on every flag, a self-check pass that re-reads the cropped region, and a visible low-confidence list. If extraction accuracy on the first six sample pairs falls below roughly 90 percent on limits and deductibles, the plan shifts to human-in-the-loop review before export rather than shipping confident errors.
Target delivery cost is under $0.40 of OCR and inference per policy pair.
4. Pricing and monetization
Founding Agency: $149 per month, up to 25 checks, PDF and CSV reports, citations and crops, sign-off stamping, share links. Per-Check Pack: $39 one-time for 5 credits, non-expiring inside 30 days, same report format. Both are card charges a principal approves without procurement.
The first revenue offer is a free completed check on a real policy pair with no signup, followed immediately by the founding plan link if the report surfaces a real discrepancy. Expected ACV is $149. Upsell path: monthly to multi-seat team plans, then to a per-check overage and an annual prepay once two months of retention are visible.
5. First-customer acquisition
The channel is direct one-to-one outreach to independent agency principals and operations managers, sourced from public state independent insurance agent association directories and local agency listings. No ads, no SEO, no marketplace.
Build a list of 80 to 100 agencies with 3 to 25 staff and a named principal. Message the exact task and the exact deliverable, never a generic AI pitch. The offer is: send one proposal and the matching issued policy, receive a finished check report in 24 hours, free, no signup. Follow up twice, at day 3 and day 7, with an anonymized sample report attached, then nudge on LinkedIn or by phone.
Secondary loop: post redacted before-and-after examples into agency-owner and CSR communities where policy checking complaints already appear.
6. 30-day execution roadmap
Days 1 to 5 build and ship the MVP. Days 6 and 7 are QA and list building. Days 8 to 14 are the first outreach wave with the first sale targeted on day 12. Days 15 to 21 convert samples to paid and ship the top requested fixes. Days 22 to 28 push volume, testimonials and reliability. Days 29 and 30 close and reconcile. Full day-by-day tasks are in the roadmap section.
7. Revenue and profit scenarios
All scenarios are projections from stated assumptions, not verified revenue. Conservative: 3 monthly agencies, $447 gross. Base: 9 agencies, $1,341 gross. Upside: 22 agencies, $3,278 gross with one refund. Hackathon profit is gross less refunds, chargebacks, processing fees and paid marketing; business profit additionally deducts hosting, tooling and OCR delivery cost. No paid marketing is assumed in any scenario because the channel is organic outreach.
8. Main risks and validation experiments
Extraction accuracy on messy carrier PDFs. Free samples satisfying buyers without conversion. E&O and data-handling objections blocking uploads. Turnaround breaking at volume. Renewal seasonality thinning demand. Each risk has a cheap test this week and explicit kill criteria in the risks section.
9. First 24-hour action plan
Claim the domain, connect a Stripe account to TrustMRR, deploy a skeleton app, assemble anonymized sample pairs, and send the first ten outreach emails before hour 24. Details in the first 24-hour plan.
Smallest sellable MVP
A two-PDF upload page that returns a citation-backed discrepancy report as PDF and CSV within 24 hours, sold as a $149/month founding plan.
Must-have features
- Two-PDF upload for proposal and issued policy, text-based and scanned
- Field extraction for policy number, dates, named insured, coverage lines, limits, deductibles and form numbers
- Normalization for currency, date formats and name word order
- Deterministic four-way diff per field: matched, mismatched, missing, extra
- Page citation plus auto-generated source crop per flag
- PDF and CSV export with reviewer initials and sign-off line
- Read-only share link for client delivery email
- Editable override and re-export without re-upload
- Stripe Checkout for $149 monthly and $39 five-check pack
- Free sample intake form with no signup required
Explicitly out of scope
- Agency management system integrations
- ISO or carrier form library
- Coverage interpretation opinions
- Commission or accounting features
- Multi-user roles, SSO, admin portal
- Claims, renewals, certificates of insurance
- Full automation without reviewer confirmation
Demo flow (what the buyer sees)
- CSR uploads the proposal PDF
- CSR uploads the matching issued policy PDF
- Pipeline parses both into a normalized field table
- Diff engine classifies each field and attaches page number and crop
- Reviewer clicks flags to verify source crops and correct one extraction
- Reviewer stamps initials and exports PDF plus CSV
- Reviewer copies the read-only share link into the client email
Technical implementation plan
Stack
Architecture
Uploads land in Supabase storage and a job row is created in Postgres. A serverless route renders pages with pdf.js, sends page images plus a per-document-family prompt to a managed vision-language OCR endpoint, and validates the response against a strict JSON schema; anything failing validation is marked low-confidence rather than guessed. A deterministic normalization and diff module compares the two field tables with no model involved, attaching page numbers and crop coordinates to every flag. The report renderer assembles cropped images and a sign-off block into PDF and CSV, stores them for the share link, and Resend emails the link to the reviewer.
Build steps
- Create Next.js repo, Supabase project and Vercel deployment with a health check
- Define field schemas for proposal, declarations page and endorsement schedule
- Build upload, page rendering and crop extraction with pdf.js
- Wire the managed OCR endpoint with schema validation and confidence flags
- Build normalization and the deterministic four-way diff engine
- Build the review UI with clickable flags and source crops
- Build PDF and CSV export with initials and share link
- Add reviewer override and re-export
- Add Stripe Checkout for monthly plan and credit pack with a credit ledger
- Add free sample intake form and email delivery
- Run six sample pairs end to end and log extraction failures
- Harden low-confidence handling and error states
AI components
- Vision-language OCR extraction of fields from proposal and policy pages
- Page classifier that decides which document family a page belongs to
- Self-check pass that re-reads the cropped region of each low-confidence field
Pricing and monetization
Founding Agency
Entry offerFor Independent P&C agency with 3 to 25 staff doing 20 or more checks per month
- Up to 25 proposal-versus-policy checks per month
- PDF and CSV discrepancy report with page citations and source crops
- Reviewer sign-off stamping
- Shareable read-only report links
Per-Check Pack
For Small or seasonal agency that wants to trial without a subscription
- 5 policy check credits
- Same report and export format
- Credits do not expire inside 30 days
Model: Monthly subscription for regular check volume, plus a small one-time credit pack for trial buyers.
First revenue offer: Send one proposal and the matching issued policy; receive a completed discrepancy report with page citations within 24 hours, free, no signup. The founding plan link follows only if the report surfaces a real discrepancy.
Expected ACV: $149
Upsell path: Per-check pack buyers move to the monthly founding plan after the second check; monthly agencies move to a multi-seat team plan and an annual prepay once two months of retention are visible.
First-customer acquisition
Direct one-to-one email outreach to independent agency principals and operations managers.
Where to find them
- Public state independent insurance agent association member directories
- Local agency listings and Google Maps results for independent agencies with 3 to 25 staff
- LinkedIn titles: Agency Principal, Operations Manager, Senior CSR at independent P&C agencies
- Agency-owner and CSR communities where policy checking complaints already appear
Exact outreach motions
- Build a list of 80 to 100 agencies with a named principal and a direct email
- Send 20 personalized emails per day naming the exact task and the exact deliverable
- Offer a free completed check on one real policy pair with 24-hour turnaround
- Follow up at day 3 and day 7 with an anonymized sample report attached
- Nudge non-openers on LinkedIn or by phone with the same one-sentence offer
- Post redacted before-and-after check examples in agency-owner communities
Message template
Hi {FirstName} — quick question about how {Agency} checks an issued policy against the proposal before it goes to the client. I built PolicyDiff: upload the proposal PDF and the issued policy PDF, get a discrepancy report with a page citation and crop behind every mismatch. Send me one pair this week and I'll return the finished report in 24 hours, free, no signup. Worth a look?30-day execution roadmap
Week 1 · Build & validate · 7 days
Accounts and deploy
- Claim domain and create Next.js repo on Vercel
- Connect Stripe account to TrustMRR
- Create Supabase project and storage buckets
- Assemble 3 anonymized sample proposal/policy pairs from public specimen documents
Deliverable: Production URL live with health check and sample pairs stored.
Upload and extraction
- Build two-PDF upload with pdf.js page rendering
- Define field schemas for proposal and policy pages
- Wire managed vision-language OCR endpoint with schema validation
Deliverable: Extracted field JSON for one sample pair.
Normalization and classification
- Normalize currency, dates and name word order
- Build four-way classifier per field
- Attach page numbers and crop coordinates to every flag
Deliverable: Diff table for one sample pair with citations.
Export artifact
- Render PDF report with crops and sign-off block
- Add CSV export
- Add read-only share link
Deliverable: Downloadable report and share link.
Payment and intake
- Add Stripe Checkout for $149 monthly and $39 pack
- Add credit ledger and free sample intake form
- Add reviewer override and re-export
Deliverable: MVP live with working checkout and intake.
Accuracy test
- Run 6 sample pairs end to end
- Log every extraction failure by field type
- Fix the two highest-frequency failures
Deliverable: Accuracy log with pass rate on limits and deductibles.
Prospect list and demo
- Build list of 80 agencies with named principals and emails
- Finalize landing copy and sample report demo link
- Draft the outreach message
Deliverable: 80-row prospect list and live sample report.
Week 2 · Launch & first sale · 7 days
First wave
- Send 20 personalized emails with the free-check offer
- Track opens and replies in a simple sheet
Deliverable: 20 emails sent and tracked.
Second wave
- Send 20 more personalized emails
- Process any inbound policy pairs
Deliverable: 40 total sent, first free sample in progress.
Free samples
- Deliver 2 free sample reports within 24 hours
- Record minutes spent per report
Deliverable: 2 delivered reports and a turnaround baseline.
Day-3 sequence
- Follow up on the day-8 batch with sample report attached
- Post a redacted example in one agency-owner community
Deliverable: Reply rate by subject line.
Convert warmest lead
- Send founding plan checkout link to the 5 warmest prospects
- Offer to run one more free check to close
Deliverable: First paying agency (target day).
Sample throughput
- Deliver all outstanding sample reports
- Ask each recipient for a 15-minute feedback call
Deliverable: All samples delivered, 2 calls booked.
Cut and add
- Cut the lowest-converting message
- Add 30 new prospects from a second state directory
Deliverable: Revised message and 110-prospect list.
Week 3 · Repeat & systematize · 7 days
Volume
- Send 25 new personalized emails
- Send LinkedIn nudges to non-openers
Deliverable: 25 sent, 3 new replies.
Sample to paid
- Call every recipient of a completed sample
- Offer the $39 pack as a low-friction entry
Deliverable: At least 2 additional paid accounts.
Top requested fix
- Ship the three most requested field or export additions
- Add low-confidence field list to the report
Deliverable: Updated report format shipped.
Community loop
- Send 25 new emails
- Post second redacted example in a CSR community
Deliverable: Second community post live.
Pack buyers
- Email all free-sample recipients who did not subscribe
- Offer expiring 5-check pack
Deliverable: Pack conversion count.
Turnaround
- Deliver all outstanding checks within 24 hours
- Measure and log average turnaround
Deliverable: Turnaround log under 24 hours.
Midpoint metrics
- Review paid count, reply rate, delivery time
- Decide whether to add a second channel
Deliverable: Midpoint metrics note.
Week 4 · Scale what works · 7 days
Third directory
- Build 30 prospects from a third state association directory
- Send 25 emails
Deliverable: 135 total prospects contacted.
Testimonials
- Ask paying agencies for a one-line quote
- Ask for one referral each
Deliverable: At least 1 usable quote and 2 referrals.
Error log
- Fix the top failure from the extraction log
- Add a retry path for failed parses
Deliverable: Reduced parse failure rate.
Subject line test
- Send 25 emails with an alternate subject line
- Compare reply rate against the control
Deliverable: Winning subject line identified.
Non-responders
- Send final follow-up to all non-responders
- Phone the 10 warmest
Deliverable: Full sequence complete on all prospects.
Refund risk
- Re-check every delivered report for missed flags
- Confirm none disputed
Deliverable: Zero disputes, QA log updated.
Forecast
- Count paid, active and churn-risk accounts
- Send renewal reminder to month-one subscribers
Deliverable: Paid account list and renewal emails sent.
Days 29-30 · Consolidate · 2 days
Final push
- Send final outreach to every warm lead
- Offer founding pricing for the last time
Deliverable: All warm leads contacted.
Profit statement
- Reconcile Stripe payouts and refunds
- Export the profit statement and write the post-contest plan
Deliverable: Final reconciled numbers and 60-day plan.
Revenue and profit scenarios
How the numbers are defined
Hackathon profit — Collected revenue − refunds − chargebacks − payment processing fees − paid marketing spend.
Business profit — Hackathon profit − operating costs (hosting, domains, tooling, AI/API delivery costs).
The contest ranks entrants on hackathon profit only. Hosting, tooling and AI/API delivery costs are not deducted by the organisers, but each sale still has to cover them — so we show both. Definitions verified on 2026-11-15 against the official rules.
Stated assumptions
- Base ARPU assumes the $149 Founding Agency monthly plan; the $39 per-check pack is a fallback and is not counted in projections.
- Processing fees assume Stripe at 2.9% plus $0.30 per charge, roughly 3.1 to 3.9 percent of collected revenue.
- No paid marketing in any scenario: acquisition is organic one-to-one outreach, so ad_spend is zero.
- Conservative assumes 3 monthly agencies from 40 sent emails; Base assumes 9 from 110 contacts; Upside assumes 22 from 135 contacts.
- Operating costs assume about $0.20 to $0.40 of OCR and inference per policy pair plus $20 to $40 of monthly hosting, domain and tooling.
- Refunds are assumed at zero in Conservative and Base and one month in Upside; chargebacks are assumed zero.
- Every scenario is a projection from stated assumptions and must not be read as verified revenue or existing customers.
Conservative
3 customers- Refunds
- $0
- Chargebacks
- $0
- Payment fees
- $16
- Paid marketing
- $0
- Hackathon profit
- $431
- Operating costs
- $35
- Business profit
- $396
Base case
9 customers- Refunds
- $0
- Chargebacks
- $0
- Payment fees
- $48
- Paid marketing
- $0
- Hackathon profit
- $1,293
- Operating costs
- $70
- Business profit
- $1,223
Optimistic
22 customers- Refunds
- $149
- Chargebacks
- $0
- Payment fees
- $118
- Paid marketing
- $0
- Hackathon profit
- $3,011
- Operating costs
- $140
- Business profit
- $2,871
Model your own numbers
Runs in your browser · nothing is saved- Collected revenue
- $1,341
- Payment fees (3.9%)
- $52
- Paid marketing
- $0
- Hackathon profit
- $1,289
- Business profit
- $1,219
Hypothetical arithmetic, not a forecast. The fee rate is a flat 3.9% assumption, refunds and chargebacks are left at zero here, and none of this is tracked data.
These scenarios are projections built from the stated assumptions. They are not verified revenue, not existing customers, and not a forecast guarantee.
Main risks and validation experiments
Risk 1Extraction accuracy on scanned or poorly tagged carrier PDFs is too low for a CSR to trust the report.
Confidence thresholds, a mandatory source crop on every flag, a re-read self-check on low-confidence fields, and a visible low-confidence list before export.
Run six sample pairs on day 6 and record a pass rate on limits and deductibles; show the worst report to two CSRs and ask if they would sign it.
Abandon or pivot if limits and deductibles accuracy stays below 90 percent after two rounds of fixes and reviewers say they must re-read the whole policy anyway.
Risk 2A free sample satisfies the buyer without conversion, because checking is already unpaid CSR work.
Make the free check the first of a recurring set, keep the report format tied to per-policy use, and ask for the second pair before the first is delivered.
Send 40 free-sample offers and measure whether recipients send a second policy pair unprompted within 7 days.
If fewer than 1 in 10 sample recipients send a second pair or pay within 10 days, the offer or the buyer is wrong.
Risk 3E&O and data-sensitivity objections block uploads to an unknown solo-founder vendor.
State plainly that files are processed by a managed parsing endpoint, deleted after export, never used for training, and that the reviewer confirms every flag.
Put the data-handling paragraph in the first outreach email and count how many prospects raise it as an objection versus ignore it.
If more than half of replying prospects refuse to upload even one redacted pair, the trust barrier is not solvable inside 30 days.
Risk 4One founder cannot both sell and honor a 24-hour free-report promise at volume.
Cap free checks at 3 per day, batch delivery into two windows, and reuse the override workflow to make corrections fast.
Log minutes per report across the first 8 reports and extrapolate to a 5-per-day load.
If a report takes more than 45 minutes of founder time, stop free checks beyond day 20 and switch to paid-only intake.
Risk 5Demand clusters around renewal dates, thinning volume inside a 30-day window.
Target agencies with mixed effective-date books and pitch the backlog of unchecked policies already in their files.
Ask the first 10 prospects how many unchecked issued policies are sitting in their files right now.
If most prospects report fewer than 5 backlog checks, the immediate-sale premise fails and outreach shifts to a different agency size band.
Hour-by-hour: the first 24 hours
Hours 1-2
- Claim the domain and create the Next.js repo
- Connect a Stripe account to TrustMRR
- Create the Supabase project
Output: Repo and payment account ready.
Hours 3-5
- Deploy a skeleton app with upload placeholder to Vercel
- Write the field schema for proposal and policy pages
Output: Live URL and a written extraction schema.
Hours 6-9
- Assemble 3 anonymized sample proposal/policy pairs from public specimen documents
- Manually extract one pair by hand to define the expected output
Output: Three test pairs and one hand-built reference report.
Hours 10-14
- Wire the managed OCR endpoint and run one pair end to end
- Print the raw extraction to inspect failures
Output: First machine extraction of a real pair.
Hours 15-20
- Write the first outreach message
- Build the first 20 prospects with named principals
Output: 20-row prospect list.
Hours 21-24
- Send the first 10 outreach emails
- Set up the reply-tracking sheet
Output: 10 emails sent, tracking live.
Full 30-day launch report
Read the full generated report
PolicyDiff — 30-Day Launch Report
Contest window: Nov 16 13:00 UTC to Dec 16 13:00 UTC, 2026. One founder. Money collected inside the window only.
1. Product opportunity and positioning
Independent P&C agencies with 3 to 25 staff run a policy check every time a carrier issues a policy. A CSR reads the issued policy against the proposal, field by field, confirming limits, deductibles, named insureds, effective dates and endorsement form numbers. It takes 30 to 90 minutes, it is the least popular hour of the day, and misses surface months later as rework and E&O exposure.
PolicyDiff performs the cross-read. The CSR uploads two PDFs, the proposal and the issued policy, and receives a discrepancy report: every field classified matched, mismatched, missing on policy or extra on policy, each flag carrying a page number and a cropped source image, exported as PDF and CSV with a reviewer-initials sign-off line, plus a read-only share link.
The positioning sentence is the sale: the CSR verifies, they do not transcribe. The buyer is the principal or operations manager who signs the card; the user is the senior CSR who gets the hour back. Alternatives are outsourced checking vendors (batch, opaque, expensive) or the status quo of unpaid CSR overtime. Nothing must be integrated, approved or migrated for the first policy to be checked.
The wedge is deliberately narrow: personal and small-commercial lines, one proposal against one policy, one exported artifact. No carrier form library, no coverage opinions, no agency management system sync. This is not a parsing API and not a comparison matrix for hardware engineers; it is an exception report for insurance retail.
Given a C-band JEV score of 61.6, the concept's strength is speed to first revenue and reachable buyers, and its weakness is operational simplicity. The plan below answers that weakness by keeping every flag reviewer-confirmed, keeping delivery stateless on managed services, and refusing custom infrastructure.
2. Smallest sellable MVP
A working upload page, an extraction pipeline, a deterministic diff, and an emailed report. The minimum sellable unit is: one proposal PDF plus one matching issued policy PDF in, one citation-backed discrepancy report out, delivered within 24 hours, priced at $149 per month for up to 25 checks.
In scope by day 5: page-level parse of text-based and scanned PDFs; extraction of policy number, dates, named insured, coverage lines, limits, deductibles and form numbers; normalization of currency, date formats and name word order; four-way classification per field; page citation and crop per flag; PDF and CSV export with initials; read-only share link; editable override and re-export; Stripe Checkout for the monthly plan and the $39 five-check pack.
Out of scope: agency management system integrations, ISO or carrier form libraries, coverage interpretation, commissions, multi-user roles, SSO, claims, renewals, certificates of insurance, and any promise of full automation.
Demo flow: upload proposal, upload policy, wait 90 seconds, review flagged fields, click any flag to see the source crop, correct one wrong extraction, stamp initials, download PDF and CSV, copy the share link into the client delivery email.
3. Technical implementation plan
Next.js on Vercel handles upload, review UI and report rendering. Supabase stores job rows in Postgres and both source PDFs plus generated crops in object storage. A managed layout-aware vision-language OCR endpoint extracts fields into a strict JSON schema per document family; schema validation failures surface as low-confidence fields rather than guesses. pdf.js renders pages for crop generation and pdf-lib assembles the final report.
Ordering: schema first, then parse, then diff, then report, then payment. A separate prompt and schema per document family (proposal, carrier declarations page, endorsement schedule) is more reliable than one general prompt. The diff engine is ordinary code, not an LLM: exact and normalized comparison, currency and unit coercion, name word-order tolerance, and a fuzzy threshold on entity names to avoid false positives.
The single largest technical risk is a scanned or poorly tagged carrier PDF returning a wrong limit or missing an endorsement, which destroys trust on the first report. Mitigation: confidence thresholding, a mandatory source crop on every flag, a self-check pass that re-reads each low-confidence crop, and a visible low-confidence list. If accuracy on the first six sample pairs falls below roughly 90 percent on limits and deductibles, the plan shifts to human-in-the-loop review before export rather than shipping confident errors. Target delivery cost is under $0.40 of OCR and inference per policy pair.
4. Pricing and monetization
Founding Agency: $149 per month, up to 25 checks, PDF and CSV reports, citations and crops, sign-off stamping, share links. Per-Check Pack: $39 one-time for 5 credits, non-expiring inside 30 days, same report format. Both are card charges a principal approves without procurement.
The first revenue offer is a free completed check on a real policy pair with no signup, followed immediately by the founding plan link if the report surfaces a real discrepancy. Expected ACV is $149. Upsell path: monthly to multi-seat team plans, then per-check overage and an annual prepay once two months of retention are visible.
5. First-customer acquisition
The channel is direct one-to-one outreach to independent agency principals and operations managers, sourced from public state independent insurance agent association directories and local agency listings. No ads, no SEO, no marketplace.
Build a list of 80 to 100 agencies with 3 to 25 staff and a named principal. Message the exact task and the exact deliverable, never a generic AI pitch. The offer is: send one proposal and the matching issued policy, receive a finished check report in 24 hours, free, no signup. Follow up at day 3 and day 7 with an anonymized sample report attached, then nudge on LinkedIn or by phone. Secondary loop: post redacted before-and-after examples into agency-owner and CSR communities where policy checking complaints already appear.
6. 30-day execution roadmap
Days 1 to 5 build and ship the MVP, with the deadline on day 5. Days 6 and 7 are QA and list building. Days 8 to 14 are the first outreach wave, with the first sale targeted on day 12. Days 15 to 21 convert samples to paid and ship the fixes buyers ask for. Days 22 to 28 push volume, testimonials and reliability. Days 29 and 30 close and reconcile. The roadmap section carries the full day-by-day tasks and deliverables; the sequencing rule is that no sales day ever happens without a shipped deliverable behind it.
7. Revenue and profit scenarios
All scenarios are projections from stated assumptions, not verified revenue. Conservative: 3 monthly agencies, $447 gross. Base: 9 agencies, $1,341 gross. Upside: 22 agencies, $3,278 gross with one refund. Hackathon profit is gross less refunds, chargebacks, processing fees and paid marketing; business profit additionally deducts hosting, tooling and OCR delivery cost. No paid marketing appears in any scenario because the channel is organic outreach.
8. Main risks and validation experiments
Extraction accuracy on messy carrier PDFs is the gating risk, tested on day 6 against six pairs with a 90 percent accuracy floor. Free samples satisfying buyers without conversion is tested by whether recipients send a second pair unprompted. E&O and data-handling objections are tested by putting the data-handling paragraph in the first email and counting objections. Turnaround breaking under volume is tested by logging founder minutes per report. Renewal seasonality is tested by asking how many unchecked policies sit in the files today. Kill criteria for each are in the risks section.
9. First 24-hour action plan
Claim the domain, connect Stripe to TrustMRR, deploy a skeleton app, assemble three anonymized sample pairs, wire the first end-to-end extraction, build 20 prospects, and send the first 10 outreach emails before hour 24. The hour blocks are in the first 24-hour plan.
This brief was generated from public data about established startups, then narrowed and scored for a solo 30-day launch. Revenue and profit figures are scenario models, not verified results, guarantees or evidence of existing customers. Validate demand yourself before you spend a month on it — and check the contest rules for what counts toward your score.
Related ideas
All ideas →Handbook Answer Cards
Turns any employee handbook PDF into 40 citation-backed Q&A cards and a policy gap report for HR consultants serving small businesses.
BoxAudit Label Forensics
BoxAudit re-checks every bounding box in a YOLO or COCO dataset against open-vocabulary detectors and returns a ranked list of mislabelled boxes plus a repaired label export, so small vision teams stop retraining on noisy labels.
StaleDoc Audit
StaleDoc Audit scans a company's existing Google Drive in about ten minutes and delivers a ranked PDF report of outdated, duplicated and orphaned documents, so an ops lead knows exactly which pages to fix or kill first.